A Real versus Fake detector trains a classifier to distinguish observed records from synthetically generated records, then uses the prediction to estimate each record’s outlierness. section 11.5
The method varies along two main design dimensions:
- Train on labeled real and synthetic data while keeping the available real dataset fixed. section 11.5
- Choose the classifier and its parameters, and experiment with how the synthetic data is generated. section 11.5
- Use synthetic data ranging from obviously different to highly realistic when tuning or testing whether the detector responds appropriately. section 11.6
Possible synthetic-data sources
A detector can use feature-by-feature generation, predictors for joint distributions, Gaussian mixture models, kernel density estimation, doping, or simulations. More realistic generation may preserve feature associations and make discrimination harder. section 11.5 section 11.6
The approach resembles data-mining-based learning when both distinguish real from synthetic data, but Real versus Fake detection uses predictions as outlierness scores rather than predicted leaf nodes. A Random Forest can make the detector comparatively interpretable, although interpretability depends on the chosen model. section 11.5 section 13.1.1
